Somewhere inside every request a researcher sent to Grok's API this month sat about 1,240 tokens of input they didn't write, couldn't see and couldn't remove. That one detail tells you more about today's AI news than any benchmark. The quiet Saturday stories all come back to the same question: what is in the pipeline that nobody told you about?
Grok and GPT disagree on 90% of the midterm ballot#
Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, asked two frontier models the same blunt question for nearly 500 races in the 2026 US midterms: who should I vote for? Name one candidate, no hedging. TechCrunch has the results, and they are about as lopsided as you'd fear:
Grok 4.7 picked the Republican in 92% of answers.
OpenAI's GPT-6.1 Sol picked the Democrat in 97%.
The two agreed in fewer than 10% of races.
In heavily Republican districts GPT still chose the Democrat in 57 of 67; in deep-blue districts Grok chose the Republican in 47 of 72.
The sourcing is just as telling. Grok cited X in 41% of its answers, a Christian conservative voter guide in 39% and Turning Point Action in 23%. GPT leaned on Wikipedia, the AP, Axios and Reddit. GPT searched more on healthcare and abortion; Grok went to immigration, taxes, crime and energy. Each model is also, in effect, choosing which issues decide the race.
Let me be fair to OpenAI. The test used the developer API with a prompt that forced a pick, and the consumer apps refused to answer. OpenAI's spokesperson said ChatGPT is "designed to be objective by default" and that this setup doesn't reflect real use. Pew says only 8% of Americans use chatbots for election information. Those are reasonable points. They are also beside the point for anyone building on these APIs, because the API is exactly where your product lives. If your civic-info bot, your newsroom tool or your "ask anything" widget calls one of these models with a direct question, this is the default behaviour you inherit.
The part that bothers me most is Grok's hidden input. Parikh called those 1,240 extra tokens potentially "hidden instructions." SpaceXAI didn't respond before publication, so we don't know what they say. In practice it means a vendor-controlled system prompt sits in front of yours, and you can't audit it. That is a supply-chain dependency, and it should be treated like one.
If you ship anything that might touch politics, run your own version of this test before November. It's cheap: a few hundred prompts, a spreadsheet, an afternoon.
Kimi K3.1 Pro's numbers are real enough; its marketing might not be#
Moonshot's Kimi K3.1 Pro is a 3-trillion-parameter post-train of July's K3, and the jumps on Artificial Analysis are big: Terminal-Bench 4.0 went from 13% to 67%, and the Intelligence Index from 44 to 55, fourth overall. API pricing is $2 input and $10 output per million tokens, the same as Claude Sonnet 5.5 and half of Opus 5.5. Artificial Analysis puts it at $2.11 per Intelligence Index task against $5.98 for Opus. Moonshot says it will release the weights under an open license on November 4, though the license terms aren't published yet.
Then there's the asterisk. An X user posted screenshots of what he described as instructions for a Kimi influencer campaign, including this line:
"Please DO NOT put on the paid partnership label"
The suggested talking points framed Kimi as an open-source Opus 5.5 at a fraction of the price. VentureBeat is careful here: the document hasn't been authenticated, it's unclear whether Moonshot, an agency or someone else wrote it, and Moonshot hadn't responded. So treat it as an allegation. But it explains why your feed has been full of glowing Kimi threads, and why I'd discount any of them that don't link to a reproducible eval.
My read: the model is probably very good at agentic and office work. Moonshot's own table has it ahead of Opus 5.5 on SWE-Marathon (51.7 vs 46.3), but well behind on Terminal-Bench Science (38.1 vs 63.3). One more thing for security teams: Artificial Analysis logged no safety refusals on its cyber evals, where Kimi scored 95% on CyberGym-E2E. Wait for the November weights and test it on your own tasks before you believe anyone's thread, mine included.
Pennsylvania wants the label on the wing-special flyer too#
Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on Thursday. It requires a label on any ad, whether image, video, audio or text, that "has been significantly modified or generated by artificial intelligence," in the same format as the ad itself. The Verge notes this covers everyone from billion-dollar TV campaigns to a restaurant's crowded flyer. Pennsylvania is one of the first states to require this.
For builders of ad and design tools, "significantly modified" is the phrase that will generate the work, because someone has to decide where retouching ends. Expect requests for provenance metadata and auto-labeling toggles in creative tools first, and for the boundary cases to land in your support queue.
The cultural version of this showed up the same day. The BBC's Blue Planet III is keeping plankton, buoys and imperfect water in its shots, because flawless footage now reads as fake. When polish becomes suspicious, the label has to do the work the image used to.
Also on the wire#
Intel asked the FCC for a waiver "out of an abundance of caution" because it couldn't tell whether the US ban on certain foreign-made robots covered the automated guided vehicles in its fabs. On October 22 it got one year. It's not an AI story in the narrow sense, but it shows how loosely drafted automation rules land on real factories: the company most associated with US chipmaking couldn't read the rule with confidence.
Every story today is about a missing label: an undisclosed system prompt, an allegedly undisclosed sponsorship, an ad that might not say it's synthetic, a regulation so vague that Intel needed a waiver just to keep its robots running. My prediction: by mid-2027 "show me the full input" becomes a standard procurement question for model APIs, the way SOC 2 reports did for SaaS. I'm adding a hidden-token check to my own eval harness this week: compare billed input tokens with what I actually sent. It's a one-line diff, and on Grok it would have caught all 1,240.